A Novel Deep Convolutional Neural Network Structure for Off-line Handwritten Digit Recognition
Yan Wen, Yi Shao, Dabo Zheng · 2019
Handwritten digit recognition is widely used nowadays because traditional methods require the extraction of feature vectors from the original data by hand. The deep convolutional neural network (DCNN) can solve this problem efficiently, and the recognition effects are better than the traditional methods. However, deep learning requires lots of data set to be trained, and a suitable learning algorithm is required to adjust a lot of parameters. We analyze the difference between convolution neural network (CNN) network and traditional neural network (NN) in the recognition of handwritten digital datasets. CNN show the obvious superiority to the digital prediction rate, and then we study the applicability of DCNN in handwritten digital datasets. A deep convolutional neural network (DCNN) based on Alex network (AlexNet) is designed and the recognition accuracy can reach to 99%. The confusing digital set are sort out, called Confuse Numbers (CN) based on MNIST. In the field of handwritten digit recognition, most of the work is focusing on the processing of the whole MNIST data set. As a supplement, the CN dataset is helpful to make the progress on handwritten digit recognition technology.